Updated: Mar 30, 2026
base modelDownload
1 variant available
Version '78 SE
Version '78 SE comes with Vpred and merges Version '78 with theWeirdChemist's QuantumDumbbell v2 topped with a distilled version of AutismMix Confetti by Autismix_anon. It does stuff. It's packed with the standard SDXL VAE so no more weird tone mismatches when inpainting.
Still generating at 1024 x 1536, using 4 CFG and 25-ish steps. Same sampler as below but using GITS for some reason because it doesn't regulate with Vpred, but I'm lazy and like it that way.
Version '78
NOTE: The packaged VAE is spoopy with inpainting so I recommend using the standard SDXL VAE for the time being. I'm currently working on finishing up my ComfyUI for Dummies node set for release and will update this ASAP.
Version '78 had me take Version 1.5 and bake in some loras:
Something Frankensteined by Me
I generate with this one using 1024 x 1536 and 1280 x 1280, 5 CFG, and 25 steps. Hires fix at 1.5x using ClearRealityV1 w/lanczos downscaler @ 0.618 denoising. Ultimate SD Upscale at 1.5x using RealESRGAN_4xplus @ 0.382 denoising, 4 tiles, 256px padding, and 32px mask blur. After some testing, ESRGAN is the only upscaler I find doesn't shift contrast or add film grain.
As before, I use deis_ode_2m with Beta57 scheduler (see: RES4LYF). Otherwise, I'd recommend using Euler with Exponential though your mileage may vary.
Positive Tags:
very aesthetic, highres, best quality, Negative Tags:
worst quality, bad quality, lowres, blurry, too many fingers, Maybe it's the higher scaling I work with, but it is a little wonky with fingers sometimes.
Version 1.5
Welcome to version 1.5, created with the aim of smoothing out issues found in 1.0 after some use. I've been using this version for about a week, and I've been much happier with it. I've replaced a couple of checkpoints in the mix-up and used the same reverse-DARE method to blend everything together.
Thanks to the new contributors:
JedpointIL by Jedas
My current technique is to generate at 768 x 1152 then hi-res fix at 2x, lanczos downscale back to 768, hi-res fix again at 2x, then 1.5x upscale with Ultimate SD Upscaler using 4x 2304 x 864 tiles. CFG 3.0 and 25 steps. It changes week to week, but I've settled in that the best two upscale models to use are either Remacri or 4x_animesharp. Everything else generates too much noise, imo.
Also back to using deis_ode_2m with Beta57 scheduler (see: RES4LYF) despite the images used in the showcase.
Version 1.0
I'm just swinging for the fences. The goal was to integrate four types of styles: toon, flat, real, and western. I took four checkpoints for each style and dare merged the lot. Then I took the same checkpoints and simple merged them seperately in the reverse order of the dare merging. Then I dare merged both chunks in favor of the original dares. So, yeah, sixteen checkpoints in all went into this. Twice.
Thanks to all the creators of:
Janku v6.0 by janxd
Illustratious v1.0 by Jedas
So far it's been minimal inpainting. Mostly hands and eyes, but I'm also sort of hot-rodding things.
Generation: 1024 x 1536, 3 CFG, 25 steps, 1.0 Denoise.
Stage 1 Hi-Res Fix: 4x-Animesharp @ 1.5 upscale, lanczos down, 3 CFG, 25 steps, 0.618 Denoise.
Stage 2 Hi-Res Fix: 4x-Animesharp @ 1.5 upscale, lanczos down, 3 CFG, 25 steps, 0.382 denoise.
25 steps on last stage is overkill, but then I got high (la da da da da da).
Sampler I use now is deis_ode_2m with Beta57 scheduler (see: RES4LYF)
In lieu of embeddings, I've been using:
Positive tags
masterwork, masterpiece, very aesthetic, bespoke, absurdres, highres, ray tracing, cinematic shadows, 8k, uhd, best quality, amazing quality, perfect composition, high detailed face, high detailed eyes, perfect eyes, absolutely gorgeous, Negative tags
worst quality, bad quality, bad anatomy, bad hands, text, artist name, signature, username, logo, watermark, copyright name, copyright symbol, artist request, I'll probably try to VPred this over the weekend.